Unicode Sinhala andPhonetic English Bi-directional ...
Embed Size (px)
Transcript of Unicode Sinhala andPhonetic English Bi-directional ...
IEEE International Conference on Industrial and Information Systems 2015 \Unicode Sinhala and Phonetic English Bi-directional
Conversion for Sinhala Speech Recognizer
M. PunchimudiyanseDepartment of Mathematics and Computer Science
The Open University of Sri LankaNawala, Sri [email protected]
Abstract-An automated speech recognizer (ASR) having alarge vocabulary is yet to be developed for the Sinhala languagebecause of the time consuming nature of gathering the trainingdata to build a language model. The dictionary and building thelanguage model require non-English text, in our case, SinhalaUnicode, to be transcribed in phonetic English text. Unlike text tospeech conversions which only require transcribing the non-English text to phonetic English text, an ASR needs correctreproduction of the original language text when the phoneticEnglish text is produced as the output of the speech recognizer.In the present research, newspaper articles are used to gather alarge set of sentences to build a language model having thousandsof words for the Sphinx ASR. We present a decoder algorithmthat produces phonetic English text from Sinhala Unicode textand an encoder algorithm that produces the correct reproductionof Unicode Sinhala text from phonetic English. For a nearphonetic tag set for Sinh ala alphabet, results indicate 100%accuracy for the decoder algorithm while for numberless text,accuracy of the encoder algorithm stands at 98.61 % for distinctphonetic English words.
Keywords=-Sinhala to Phonetic English, Phonetic English toSinhala, Sinh ala ASR, Sinh ala Phonetic Tag set, Sphinx
Commercial and open source products to automaticallyrecognize spoken English is available in the market today. Interms of performance and the accuracy, ASR systems that arebased on acoustic and language models utilize hidden markovmodel (HMM) or Gaussian mixture models (GMM) forrecognition. Most popular such open source ASR systemsinclude Sphinx, HTK and Julius [1-3]. HMM based ASRsystems typically have two phases, namely, the training phaseand the recognition phase . Onetime training phase is usedin which a single or multiple user's voice is utilized to train aproper speech model called an acoustic model. In therecognition phase, the speech model built in the previousphase is recurrently used to recognize spoken words of thespeaker.
The process of training a speech model requires acollection of acoustic data, a recorded voice sample of aperson, and the correct phonetic text transcription of that voicesample. Adopting this technique for systems built for non-English languages require a production of non-English wordsin phonetic English text for the speech trainer. In order to
R. G. N. MeegamaDepartment of Computer ScienceUniversity of Sri Jayawardenapura
Gangodawila, Sri [email protected]
build the phonetic text transcription from the original Unicodetext, a proper phonetic tag per character is necessary.
Sinhala Unicode text to phonetic English tagging(syllabification) is not a straightforward task because Sinhalacharacters have several modifiers that are typed after theactual base character but appears before, after, top and thebottom of the actual base character. In addition, a nonprintable zero-width joiner (Unicode U+200D) is insertedbetween the characters to convert certain charactercombinations to one of the modifiers named yansaya,rakaranshaya or repaya which are not present in SinhalaUnicode.
For an ASR application to work, both Sinhala Unicode tophonetic text conversion and proper reproduction of SinhalaUnicode from phonetic English output of the speechrecognizer is necessary. Phonetic tags for Sinhala alphabet hasto be searched and replaced in a specific order to preventinvalid construction of words from phonetic English. Forexample, if phonetic tag for the characters cy is i and C5 is iiwith search order i, ii is used to encode a phonetic Englishword giithaya (meaning song), the word is constructed as (en +cy + cy + 25) + ce) a3CY25)~ which is misspeJled. Correct spelling ofthe word ~25)~ will be constructed if the search order ii,i isfollowed.
The ASR applications use N-gram based techniques toconstruct word combinations to build a language model inphonetic English. Therefore, it is necessary for the researchersto identify suitable phonetic tag set to develop a Sinhala tophonetic text decoder algorithm and phonetic English toSinhala Unicode encoding algorithm.
II. BACKGROUND AND RELATED WORK
Representation of Sinhala characters in digital formcommenced in early 1980s to display Sinhala subtitles intelevision. The Wijesekara keyboard used in the type writersare extended to Sinhala fonts and keyboard drivers used in thecomputers. The Sinhala alphabet is first standardized by theSri Lanka standards (SLS) institute as Sinhala Character Codefor Information Interchange (SLS 1134: 1996) andsubsequently revised in 2004 and 2011 (SLS 1134:2004, SLS1134:2011) . The Sinhala alphabet representation inUnicode is first added to the Unicode standard version 3 basedon the ISO/lEC 10646 .